CN105930797A

CN105930797A - Face verification method and device

Info

Publication number: CN105930797A
Application number: CN201610250901.5A
Authority: CN
Inventors: 汪铖杰; 李季檩; 梁亦聪
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd; Tencent Cloud Computing Beijing Co Ltd
Priority date: 2016-04-21
Filing date: 2016-04-21
Publication date: 2016-09-07
Anticipated expiration: 2036-04-21
Also published as: EP3370188A4; KR20180077239A; JP6656376B2; EP3370188B1; JP2019500687A; KR102045978B1; WO2017181923A1; CN105930797B; EP3370188A1

Abstract

The invention discloses a face verification method and device. The method comprises: obtaining a textured face image and an initial face image needing verification; using a preset texture removing model to perform texture removing operation of the textured face image to obtain a de-textured face image; extracting features of the de-textured face image by means of a preset de-textured face verification model to obtain de-textured face features, and extracting features from the initial face image by means of the preset de-textured face verification model to obtain initial face features; and based on the de-textured face features and the initial face features, verifying the de-textured face image and the initial face image. According to the de-texture model and the de-textured face verification model trained in advance, the method conducts automatic processing and analysis of the textured face image and initial face image needing verification, with no need of manual operation by professionals. The method can automatically determine whether the textured face image and the initial face image are of the same person and improve the efficiency and accuracy of face verification.

Description

A kind of face verification method and device

Technical field

The invention belongs to communication technical field, particularly relate to a kind of face verification method and device.

Background technology

Along with developing rapidly of Internet technology, ensure that user account safety also seems by face verification Particularly important.Face verification, is a branch of field of face identification, and use face verification algorithm can entirely certainly Move and two human face photos are verified, it determines whether for same people.This mode can be used for the Internet finance Etc. the user's face identity verification under multiple scenes.

At present, in order to protect citizen privacy safety, reticulate pattern watermark can be added when photo unofficial biography, thus obtain Reticulate pattern photo, such as, the certificate photograph such as identity card face photograph, social security card face photograph, pass face photograph. When reticulate pattern photo is carried out face verification, professional is needed to remove net from reticulate pattern photo by Denoising Algorithm Stricture of vagina, carries out reparation etc. to the photo after removing reticulate pattern the most again, finally could verify this reticulate pattern photo.

To in the research of prior art and practice process, it was found by the inventors of the present invention that due to existing removal The image procossing mode of reticulate pattern needs manually to carry out, and requires higher, therefore to the professional technique of operator The inefficient of face verification, the problem that face verification accuracy is the highest can be caused.

Summary of the invention

It is an object of the invention to provide a kind of face verification method and device, it is intended to improve the effect of face verification Rate and accuracy.

For solving above-mentioned technical problem, embodiment of the present invention offer techniques below scheme:

A kind of face verification method, including:

Reticulate pattern facial image that acquisition need to be verified and Initial Face image；

Utilize the descreening model preset that described reticulate pattern facial image is carried out descreening operation, obtain descreening Facial image；

By default descreening face verification model, described descreening facial image is carried out feature extraction, To descreening face characteristic, and by described default descreening face verification model to described Initial Face figure As carrying out feature extraction, obtain Initial Face feature；

Based on described descreening face characteristic and described Initial Face feature, to corresponding descreening facial image Face verification is carried out with Initial Face image.

For solving above-mentioned technical problem, the embodiment of the present invention also provides for techniques below scheme:

A kind of face verification device, including:

Acquiring unit, for obtaining the reticulate pattern facial image and Initial Face image that need to verify；

First descreening unit, for utilizing default descreening model to remove described reticulate pattern facial image Reticulate pattern operates, and obtains descreening facial image；

Feature extraction unit, is used for by default descreening face verification model described descreening face figure As carrying out feature extraction, obtain descreening face characteristic, and by described default descreening face verification mould Type carries out feature extraction to described Initial Face image, obtains Initial Face feature；

Authentication unit, for based on described descreening face characteristic and described Initial Face feature, to accordingly Descreening facial image and Initial Face image carry out face verification.

Relative to prior art, the embodiment of the present invention, first, utilize the descreening model preset automatically to needing Reticulate pattern facial image to be verified carries out descreening operation, obtains descreening facial image；Then pass through then, The descreening face verification model preset carries out feature respectively to descreening facial image and Initial Face image Extract, obtain descreening face characteristic and Initial Face feature；It is finally based on descreening face characteristic with just Beginning face characteristic, carries out face verification to corresponding descreening facial image and Initial Face image；I.e. this Needs, according to the good descreening model of training in advance and descreening face verification model, are verified by bright embodiment Reticulate pattern facial image and Initial Face image carry out automatic Treatment Analysis, it is not necessary to professional is artificial Operation, can automatically determine reticulate pattern facial image and whether Initial Face image is same people, improve The efficiency of face verification and accuracy.

Accompanying drawing explanation

Below in conjunction with the accompanying drawings, by the detailed description of the invention of the present invention is described in detail, the skill of the present invention will be made Art scheme and other beneficial effect are apparent.

Fig. 1 a is the scene schematic diagram of the face verification method that the embodiment of the present invention provides；

Fig. 1 b is the schematic flow sheet of the face verification method that first embodiment of the invention provides；

Fig. 1 c is the schematic diagram of reticulate pattern facial image in the face verification method that first embodiment of the invention provides；

The schematic flow sheet of the face verification method that Fig. 2 a provides for second embodiment of the invention；

Detailed process schematic diagram in the face verification method that Fig. 2 b to Fig. 2 e provides for second embodiment of the invention；

The structural representation of the face verification device that Fig. 3 a provides for third embodiment of the invention；

Another structural representation of the face verification device that Fig. 3 b provides for third embodiment of the invention；

The structural representation of the server that Fig. 4 provides for fourth embodiment of the invention.

Detailed description of the invention

Refer to graphic, the most identical element numbers represents identical assembly, and the principle of the present invention is with reality The computing environment that Shi Yi is suitable illustrates.The following description is concrete based on the illustrated present invention Embodiment, it is not construed as limiting other specific embodiment that the present invention is the most detailed herein.

In the following description, the specific embodiment of the present invention will be with reference to by performed by one or multi-section computer Step and symbol illustrate, unless otherwise stating clearly.Therefore, these steps and operation will have mention for several times by Computer performs, and computer as referred to herein performs to include by representing with the data in a structuring pattern The operation of computer processing unit of electronic signal.This operation is changed these data or is maintained at this calculating Position in the memory system of machine, it is reconfigurable or other with the side known to the tester of this area Formula changes the running of this computer.The data structure that these data are maintained is the provider location of this internal memory, its Have by particular characteristics defined in this data form.But, the principle of the invention illustrates with above-mentioned word, It is not represented as a kind of restriction, and this area tester will appreciate that plurality of step and the behaviour of the following stated Also may be implemented in the middle of hardware.

Term as used herein " module " can regard the software object as performing on this arithmetic system as.This Different assemblies, module, engine and service described in literary composition can be regarded as the objective for implementation on this arithmetic system. And device and method as herein described is preferably implemented in the way of software, the most also can be enterprising at hardware Row is implemented, all within scope.

The embodiment of the present invention provides a kind of face verification method and device.

See Fig. 1 a, the scene schematic diagram of the face verification method that this figure is provided by the embodiment of the present invention, should Scene can include face verification device, is mainly used in obtaining the reticulate pattern facial image that need to verify and initial people Face image, carries out descreening operation first with default descreening model to described reticulate pattern facial image, To descreening facial image；Then, by default descreening face verification model to described descreening face Image carries out feature extraction, obtains descreening face characteristic, and by described default descreening face verification Model carries out feature extraction to described Initial Face image, obtains Initial Face feature；Finally, based on described Descreening face characteristic and described Initial Face feature, to corresponding descreening facial image and Initial Face figure As carrying out face verification, etc..

Additionally, this scene can also include memorizer, be mainly used in storage preset descreening model and The descreening face verification model preset, and be supplied to face verification device and call, to carry out descreening respectively Operation and feature extraction etc..Certainly, this scene can also include user terminal, be used for receiving user defeated The need entered carry out reticulate pattern facial image and the Initial Face image of face verification, as identity card facial image with And auto heterodyne facial image, and send to face verification device and carry out face verification.

To be described in detail respectively below.

First embodiment

In the present embodiment, will be described from the angle of face verification device, this face verification device is concrete Can be integrated in the network equipments such as server.

A kind of face verification method, including: obtain the reticulate pattern facial image and Initial Face image that need to verify； Utilize the descreening model preset that described reticulate pattern facial image is carried out descreening operation, obtain descreening face Image；By default descreening face verification model, described descreening facial image is carried out feature extraction, Obtain descreening face characteristic, and by described default descreening face verification model to described Initial Face Image carries out feature extraction, obtains Initial Face feature；Based on described descreening face characteristic and described initially Face characteristic, carries out face verification to corresponding descreening facial image and Initial Face image.

Refer to the flow process signal that Fig. 1 b, Fig. 1 b is the face verification method that first embodiment of the invention provides Figure.Described method includes:

In step S101, obtain the reticulate pattern facial image and Initial Face image that need to verify.

In step s 102, utilize the descreening model preset that described reticulate pattern facial image is carried out descreening Operation, obtains descreening facial image.

Wherein, described step S101 and step S102 can particularly as follows:

Such as, face verification device first obtains the reticulate pattern facial image and Initial Face image that need to verify, so Whether the face in this reticulate pattern facial image of rear differentiation and Initial Face image is same people, to complete face Checking.

Wherein, in the embodiment of the present invention, reticulate pattern facial image refers at the human face photo that with the addition of reticulate pattern watermark, Such as the certificate photograph such as identity card face photograph, social security card face photograph, pass face photograph, can in the lump with reference to Fig. 1 c, Rough schematic for reticulate pattern facial image；Initial Face image refers to the human face photo without reticulate pattern watermark, as with The human face photo etc. of family oneself shooting.

It is understood that in the embodiment of the present invention, obtaining the reticulate pattern facial image that need to verify and initial Before facial image, need to generate the descreening model preset, wherein may particularly include following steps:

(1) collect first kind facial image, and build the first human face data according to described first kind facial image Collection.

(2) based on default reticulate pattern watermark, described first face data set is synthesized, obtains the first reticulate pattern Human face data collection.

(3) described first face data set and described first reticulate pattern human face data collection is utilized to carry out descreening convolution Neural metwork training, generates descreening model.

(4) described descreening model is defined as the descreening model preset.

Such as, face verification device collects multiple human face photos, and these multiple human face photos are defined as first Class facial image, builds the first face data set according to first kind facial image, by this first face data set It is recorded in preset memory, in order to call.

Further, face verification device utilizes the reticulate pattern watermark preset, to the in the first face data set One class facial image synthesizes, and i.e. adds reticulate pattern watermark on first kind facial image, thus obtains first Reticulate pattern human face data collection.Then, face verification device utilizes this first face data set and the first reticulate pattern face Data set carries out descreening convolutional neural networks training, thus generates descreening model.

Wherein, descreening convolutional neural networks is descreening CNN network, it may include convolutional layer (convolution), normalization layer (batch_normalization), warp lamination (deconvolution) Deng.In the embodiment of the present invention, descreening convolutional neural networks training optimization aim be so that descreening after Face shines the absolute value of the margin of image element with corresponding original face photograph and is minimum.

It is understood that concrete convolutional layer, normalization layer, the definition of warp lamination and descreening convolution The training method of neutral net is referred to existing deep neural network training framework and goes to realize, and does not makees Specifically describe.

It should be noted that term " first " in the embodiment of the present invention, " second ", " the 3rd " etc. The photo that is used only for distinguishing each type in description etc. is described rather than for representing each type Logical relation between photo or ordering relation etc..

In step s 103, by default descreening face verification model to described descreening facial image Carry out feature extraction, obtain descreening face characteristic, and by described default descreening face verification model Described Initial Face image is carried out feature extraction, obtains Initial Face feature.

Such as, face verification device can previously generate descreening face verification model, and is saved in memorizer In, the reticulate pattern facial image that need to verify is being carried out descreening operation, after obtaining descreening facial image, Call default descreening face verification model, respectively descreening facial image and Initial Face image are carried out Feature extraction, thus obtain descreening face characteristic and Initial Face feature.

Wherein, in the embodiment of the present invention, feature extraction is a concept in computer vision and image procossing. It refers to use computer to extract image information, determines whether the point of each image belongs to a characteristics of image. The result of feature extraction is that the point on image is divided into different subsets, point that these subsets tend to belong to isolate, Continuous print curve or continuous print region, in order to face verification device contrasts according to the result of feature extraction And differentiation.

It is understood that in the embodiment of the present invention, the reticulate pattern face that face verification device need to be verified in acquisition Before image and Initial Face image, need to generate the first face verification model, thus the first according to this The descreening face verification model that face checking model generation is preset, wherein generating the first face verification model can have Body comprises the steps:

A () collects Equations of The Second Kind facial image, and build the second human face data according to described Equations of The Second Kind facial image Collection；

B () obtains the 3rd class facial image that described second human face data collection is corresponding, and according to described 3rd class Facial image builds the 3rd human face data collection；

C () utilizes described second human face data collection and described 3rd human face data collection to carry out face verification convolution god Through network training, generate the first face verification model.

Such as, face verification device collects multiple human face photos, and these multiple human face photos are defined as second Class facial image, builds the second human face data collection according to Equations of The Second Kind facial image, by this second human face data collection It is recorded in preset memory, in order to call.

It should be noted that in the embodiment of the present invention, described first kind facial image and described Equations of The Second Kind face Image can be all identity card human face photo, but, described first kind facial image and described Equations of The Second Kind face Image all differs, e.g., if identity card human face photo A belongs to first kind facial image, then identity witness Face picture A is just not belonging to Equations of The Second Kind facial image.It is to say, the first face data set and the second face number Do not repeat according to the facial image concentrated.

Further, step (b) obtains the 3rd class facial image that the second human face data collection is corresponding, permissible Including: obtain identity card human face photo (the Equations of The Second Kind facial image that the i.e. second human face data is concentrated) correspondence Auto heterodyne human face photo (the i.e. the 3rd class facial image), and build auto heterodyne face according to described auto heterodyne human face photo The data set of photo.

Subsequently, face verification device utilizes the second human face data collection and the 3rd human face data collection to carry out face verification Convolutional neural networks is trained, thus generates the first face verification model, and according to this first face verification model Generate preset descreening face verification model, wherein generate descreening face verification model may particularly include as Lower step:

D described second human face data collection, based on default reticulate pattern watermark, is synthesized, obtains the second reticulate pattern by () Human face data collection；

E every image that () utilizes described descreening model to concentrate described second reticulate pattern human face data goes Reticulate pattern operates, and generates descreening human face data collection；

(f) according to described descreening human face data collection and described 3rd human face data collection, to described first face Checking model carries out network fine setting, generates the second face verification model；

G described second face verification model is defined as the descreening face verification model preset by ().

Such as, face verification device is based on default reticulate pattern watermark, the second human face data obtaining step (a) The Equations of The Second Kind facial image concentrated adds reticulate pattern watermark, thus obtains the second reticulate pattern human face data collection；Thereafter, Every image that this second reticulate pattern human face data is concentrated by the descreening model utilizing step (3) to obtain goes Reticulate pattern operates；Finally, according to descreening human face data collection and the 3rd human face data collection, step (c) is obtained The first face verification model carry out network fine setting, thus generate the second face verification model, that i.e. presets goes Reticulate pattern face verification model.

In this embodiment, the first face verification model can be regarded as " Equations of The Second Kind face the-the three class face " Checking model, as " identity card face-auto heterodyne face " verifies model；Second face verification model can be regarded as The checking model of " descreening face the-the three class face ", such as " descreening identity card face-auto heterodyne face " Checking model.

It is understood that face verification convolutional neural networks can be with specific reference to existing deep neural network people Face verification algorithm realizes, and the mode of the training method of network and network fine setting refers to existing degree of depth god Go to realize through network training framework, be not specifically described herein.

In step S104, based on described descreening face characteristic and described Initial Face feature, to accordingly Descreening facial image and Initial Face image carry out face verification.

Such as, face verification device is based on described descreening face characteristic and described Initial Face feature, to phase The descreening facial image answered and Initial Face image carry out the mode of face verification and may particularly include:

S1041, the similarity between described descreening face characteristic and described Initial Face feature is counted Calculate, obtain similar value；

S1042, based on described similar value, it is judged that whether corresponding descreening facial image and Initial Face image Belong to same people；

If S1043 judges to belong to same people, it is determined that face verification is passed through；

If S1044 judges to be not belonging to same people, it is determined that face verification is not passed through.

Such as, face verification device to state descreening face characteristic and described Initial Face feature carry out feature away from From calculating, to obtain the similarity between described descreening face characteristic and described Initial Face feature.

Further, based on described similar value, it is judged that corresponding descreening facial image and Initial Face image Whether belong to same people (i.e. step S1042) can specifically include:

(1) if described similar value exceedes predetermined threshold value, it is determined that corresponding descreening facial image is with initial Facial image belongs to same people；

(2) if described similar value is less than predetermined threshold value, it is determined that corresponding descreening facial image is with just Beginning facial image is not belonging to same people.

In some embodiments, to the phase between described descreening face characteristic and described Initial Face feature Like degree carry out calculating (i.e. step S1041) may include that utilize Euclidean distance algorithm, COS distance algorithm, Any one in associating bayesian algorithm and metric learning algorithm, to described descreening face characteristic and described Similarity between Initial Face feature calculates.

Wherein, if similar value exceedes predetermined threshold value, it is determined that corresponding descreening facial image and initial people Face image belongs to same people, i.e. face verification device and determines that face verification is passed through, if similar value is less than pre- If threshold value, it is determined that corresponding descreening facial image is not belonging to same people, i.e. face with Initial Face image Checking device determines that face verification is not passed through.

From the foregoing, the face verification method that the embodiment of the present invention provides, first, utilize that presets to remove net Stricture of vagina model carries out descreening operation to the reticulate pattern facial image needing checking automatically, obtains descreening facial image； Then, then by default descreening face verification model to descreening facial image and Initial Face image Carry out feature extraction respectively, obtain descreening face characteristic and Initial Face feature；It is finally based on descreening Face characteristic and Initial Face feature, carry out face to corresponding descreening facial image and Initial Face image Checking；I.e. the embodiment of the present invention is according to the good descreening model of training in advance and descreening face verification model, The reticulate pattern facial image and Initial Face image needing checking is carried out automatic Treatment Analysis, it is not necessary to specially Industry personnel's manual operation, can automatically determine reticulate pattern facial image and whether Initial Face image is same People, improves efficiency and the accuracy of face verification.

Second embodiment

According to the method described by first embodiment, below citing is described in further detail.

In the present embodiment, as a example by the auto heterodyne human face photo of reticulate pattern identity card human face photo and user, to net The proof procedure of part of tatooing card human face photo and auto heterodyne human face photo is described in detail.

Refer to the flow process signal of the face verification method that Fig. 2 a, Fig. 2 a provides for second embodiment of the invention Figure.Described method flow specifically includes that S21, generates descreening model；S22, generation identity card face- The checking model of auto heterodyne face；S23, the checking model of generation descreening identity card face-auto heterodyne face；S24、 Reticulate pattern identity card facial image and auto heterodyne facial image are verified.Can be specific as follows:

S21, generation descreening model.

The schematic flow sheet of descreening model can be generated for the present invention in the lump with reference to Fig. 2 b, including:

Step S211: build identity card face according to data set 1.

Such as, face verification device collects identity card facial image, and builds according to this identity card facial image Identity card face is according to data set 1.

Step S212: generate reticulate pattern identity card face and shine.

Step S213: build reticulate pattern identity card face according to data set 1.

Such as, face verification device randomly selects multiple reticulate pattern watermark, to identity card face according in data set 1 Each face shine into row reticulate pattern watermark synthesis, thus obtain the reticulate pattern identity card face photograph of multiple correspondences, For completing structure reticulate pattern identity card face according to data set 1.

It is understood that the building-up process of reticulate pattern watermark can be specific as follows: make multiple reticulate pattern watermark in advance, Image pixel value is shone by directly reticulate pattern partial pixel value being replaced identity card face, or by watermark pixel value The mode merged by a certain percentage according to image pixel value with identity card face completes the conjunction of reticulate pattern identity card face photograph Become, etc., it is not especially limited herein.

Step S214: descreening convolutional neural networks is trained, generates descreening model.

Such as, face verification device utilizes identity card face according to data set 1 and corresponding reticulate pattern identity card face Descreening CNN (convolutional Neural) network training is carried out, it is thus achieved that descreening model according to data set 1.Wherein, Descreening CNN network includes convolutional layer, normalization layer, warp lamination etc., and training optimization aim is for removing net Identity card face after stricture of vagina shines the absolute value of the margin of image element with corresponding original identity card face photograph and is Little.

S22, the checking model of generation identity card face-auto heterodyne face.

Can be in the lump with reference to Fig. 2 c, for the flow process signal of the checking model of identity card face-auto heterodyne face of the present invention Figure, including:

Step S221: build identity card face according to data set 2.

Such as, face verification device collects identity card facial image, and builds according to this identity card facial image Identity card face is according to data set 2.

It should be noted that in the embodiment of the present invention, described identity card face is according to data set 1 and described identity The picture that witness's face takes in data set 2 all differs, e.g., if identity card human face photo A belongs to identity card Face is according to data set 1, then identity card human face photo A is just not belonging to identity card face according to data set 2.The most just Being to say, identity card face does not repeats according to the facial image in data set 1 with identity card face according to data set 1.

Step S222: build auto heterodyne face according to data set 1.

Such as, face verification device collection identity card face is right according to identity institute according to identity card face in data set 2 Answer the auto heterodyne face photograph of people, and according to auto heterodyne face according to building auto heterodyne face according to data set 1.

Step S223: face verification convolutional neural networks is trained, generates testing of identity card face-auto heterodyne face Model of a syndrome.

Such as, face verification device utilizes identity card face to enter according to data set 1 according to data set 2 and auto heterodyne face Row face verification CNN network training, it is thus achieved that " identity card face-auto heterodyne face " checking model.

S23, the checking model of generation descreening identity card face-auto heterodyne face.

The checking model of descreening identity card face-auto heterodyne face can be generated for the present invention in the lump with reference to Fig. 2 d Schematic flow sheet, including:

Step S231: build reticulate pattern identity card face according to data set 2.

Such as, face verification device uses identity card face to carry out reticulate pattern identity card face according to raw according to data set 2 Become, and build reticulate pattern identity card face according to data set 2.Wherein, reticulate pattern identity card face is according in data set 2 Reticulate pattern identity card face is according to the rapid S212 of course synchronization generated and step S213.

Step S232: according to data set 2, reticulate pattern identity card face is carried out descreening operation, generates descreening body Part witness's face is according to data set.

Such as, the descreening model obtained during face verification device uses step S214 is to reticulate pattern identity witness Face carries out descreening operation according to the every pictures in data set 2, generates descreening identity card face according to data set.

Step S233: the checking model of identity card face-auto heterodyne face is carried out network fine setting, generates and remove net Tatoo a part checking model for witness's face-auto heterodyne face.

Such as, face verification device uses descreening identity card face according to data set and human face data collection of certainly taking pictures " identity card face-auto heterodyne face " checking model obtained in 1 pair of step S223 carries out network fine setting, obtains Obtain " descreening identity card face-auto heterodyne face " checking model.

It is understood that face verification CNN network can be with specific reference to existing deep neural network face Verification algorithm realizes, and it is neural that the mode of the training method of network and network fine setting refers to the existing degree of depth Network training framework goes to realize, and is not specifically described herein.

S24, reticulate pattern identity card facial image and auto heterodyne facial image are verified.

Reticulate pattern identity card facial image and auto heterodyne facial image can be tested for the present invention in the lump with reference to Fig. 2 e The schematic flow sheet of card, including:

Step S241: reticulate pattern identity card face is shone into the operation of row descreening, generates descreening identity card face According to.

Such as, the descreening model obtained during face verification device uses step S214 is thrown the net to the one of input Part witness's face of tatooing shines into the operation of row descreening, it is thus achieved that descreening identity card face shines.

Step S242: descreening identity card face is shone into row feature extraction, obtains descreening identity card face Feature.

Step S243: auto heterodyne face is shone into row feature extraction, obtains auto heterodyne face characteristic.

Obtain during further, face verification device uses step S233 " descreening identity card face-from Clap face " verify that descreening identity card face is shone into row feature extraction by model, it is thus achieved that descreening identity witness Face feature；Use " descreening identity card face-auto heterodyne face " checking model pair obtained in step S233 Auto heterodyne face shines into row feature extraction, it is thus achieved that auto heterodyne face characteristic.

Step S244: descreening identity card face characteristic and auto heterodyne face characteristic are carried out characteristic distance meter Calculate.

Such as, face verification device calculates the phase between descreening identity card face characteristic and auto heterodyne face characteristic Like degree.Wherein, the calculating of similarity can use Euclidean distance, cos distance or use the most senior Associating bayes method or metric learning method etc..

Step S245: obtain face verification result according to result of calculation.

Such as, face verification device by by Similarity Measure result compared with predetermined threshold value, it is thus achieved that final The result, it is determined that reticulate pattern identity card face according to and auto heterodyne face according to whether belonging to same people.

Can be concrete, if similarity exceedes predetermined threshold value, it is determined that corresponding reticulate pattern identity card face according to and Auto heterodyne face photograph belongs to same people, i.e. face verification device and determines that face verification is passed through, if similarity does not surpasses Cross predetermined threshold value, it is determined that corresponding reticulate pattern identity card face shines and auto heterodyne face is according to being not belonging to same people, i.e. Face verification device determines that face verification is not passed through.

It should be noted that the present embodiment generates descreening model can add image by conventional images segmentation The mode repaired replaces realizing, and the checking model generating identity card face-auto heterodyne face can be directly by existing High dimensional feature method realizes, and the checking model generating descreening identity card face-auto heterodyne face can use The method of existing transfer learning realizes, and is not specifically described herein.

It is to say, the embodiment of the present invention is by pretreatment reticulate pattern identity card face photograph, and after using pretreatment Reticulate pattern identity card face photograph and auto heterodyne face shine into the training method of row face verification model, and obtain height with this " reticulate pattern identity card face photograph-auto heterodyne face shines " verification algorithm of accuracy rate.Another it is contemplated that this reality Execute identity card face in example and take the picture of types such as can also taking for passport face, social security card face takes, herein Citing does not constitute limitation of the invention.

3rd embodiment

For ease of preferably implementing the face verification method that the embodiment of the present invention provides, the embodiment of the present invention also carries For a kind of device based on above-mentioned face verification method.The wherein implication of noun and the method for above-mentioned face verification In identical, implement the explanation that details is referred in embodiment of the method.

Refer to the structural representation of the face verification device that Fig. 3 a, Fig. 3 a provides for the embodiment of the present invention, Wherein can include acquiring unit the 301, first descreening unit 302, feature extraction unit 303 and checking Unit 304.

Described acquiring unit 301, for obtaining the reticulate pattern facial image and Initial Face image that need to verify； Described first descreening unit 302, for utilizing default descreening model to enter described reticulate pattern facial image Row descreening operates, and obtains descreening facial image.

Wherein, in the embodiment of the present invention, reticulate pattern facial image refers at the human face photo that with the addition of reticulate pattern watermark, Such as the certificate photograph such as identity card face photograph, social security card face photograph, pass face photograph；Initial Face image refers to The human face photo of watermark without reticulate pattern, such as the human face photo etc. of user oneself shooting.

Described feature extraction unit 303, for removing net by default descreening face verification model to described Stricture of vagina facial image carries out feature extraction, obtains descreening face characteristic, and by described default descreening people Face checking model carries out feature extraction to described Initial Face image, obtains Initial Face feature；Described checking Unit 304, for based on described descreening face characteristic and described Initial Face feature, to removing net accordingly Stricture of vagina facial image and Initial Face image carry out face verification.

It is understood that in the embodiment of the present invention, obtaining the reticulate pattern facial image that need to verify and initial Before facial image, need to generate the descreening model preset, please also refer to Fig. 3 b, fill for face verification Another structural representation put, described face verification device can also include:

First construction unit the 3051, first synthesis unit the 3052, first training unit 3053 and first is true Cell 3054, for generating default descreening model, can be concrete:

Described first construction unit 3051, is used for collecting first kind facial image, and according to described first kind people Face picture construction the first face data set；Described first synthesis unit 3052, is used for based on default reticulate pattern watermark, Described first face data set is synthesized, obtains the first reticulate pattern human face data collection.

Described first training unit 3053, is used for utilizing described first face data set and described first reticulate pattern people Face data set carries out descreening convolutional neural networks training, generates descreening model；Described first determines unit 3054, for described descreening model being defined as the descreening model preset.

Wherein, descreening convolutional neural networks is descreening CNN network, it may include convolutional layer, normalizing Change layer, warp lamination etc..In the embodiment of the present invention, the optimization aim of descreening convolutional neural networks training is Make the face after descreening according to the absolute value of the margin of image element with corresponding original face photograph with for minimum.

It is understood that in the embodiment of the present invention, the reticulate pattern face that face verification device need to be verified in acquisition Before image and Initial Face image, need to generate the first face verification model, thus the first according to this The descreening face verification model that face checking model generation is preset, refers to Fig. 3 b, described face verification device Can also include: the second construction unit the 3061, the 3rd construction unit 3062 and the second training unit 3063, For generating the first face verification model, can be concrete:

Described second construction unit 3061, is used for collecting Equations of The Second Kind facial image, and according to described Equations of The Second Kind people Face picture construction the second human face data collection.

Described 3rd construction unit 3062, for obtaining the 3rd class face that described second human face data collection is corresponding Image, and build the 3rd human face data collection according to described 3rd class facial image；Described second training unit 3063, it is used for utilizing described second human face data collection and described 3rd human face data collection to carry out face verification convolution Neural metwork training, generates the first face verification model.

Further, the 3rd class facial image that middle acquisition the second human face data collection is corresponding, may include that and obtain Take the auto heterodyne face that identity card human face photo (the Equations of The Second Kind facial image that the i.e. second human face data is concentrated) is corresponding Photo (the i.e. the 3rd class facial image), and the number of auto heterodyne human face photo is built according to described auto heterodyne human face photo According to collection.

Subsequently, face verification device utilizes the second human face data collection and the 3rd human face data collection to carry out face verification Convolutional neural networks is trained, thus generates the first face verification model, and according to this first face verification model Generating the descreening face verification model preset, refer to Fig. 3 b, described face verification device can also include: Second synthesis unit the 3071, second descreening unit 3072, fine-adjusting unit 3073 and second determine unit 3074, for generating default descreening face verification model, can be concrete:

Described second synthesis unit 3071, for based on default reticulate pattern watermark, to described second human face data collection Synthesize, obtain the second reticulate pattern human face data collection；Described second descreening unit 3072, is used for utilizing institute State descreening model and every image of described second reticulate pattern human face data concentration is carried out descreening operation, generate Descreening human face data collection.

Described fine-adjusting unit 3073, for according to described descreening human face data collection and described 3rd human face data Collection, carries out network fine setting to described first face verification model, generates the second face verification model；Described Two determine unit 3074, for described second face verification model is defined as the descreening face verification preset Model.

Such as, face verification device based on default reticulate pattern watermark, that the second human face data obtained is concentrated Two class facial images add reticulate pattern watermark, thus obtain the second reticulate pattern human face data collection；Thereafter, utilization obtains Descreening model every image that this second reticulate pattern human face data is concentrated carry out descreening operation；Finally, According to descreening human face data collection and the 3rd human face data collection, the first face verification model obtained is carried out net Network is finely tuned, thus generates the second face verification model, the descreening face verification model i.e. preset.

Can be concrete, described authentication unit 304 may include that

Computation subunit 3041, between described descreening face characteristic and described Initial Face feature Similarity calculates, and obtains similar value；Judgment sub-unit 3042, for based on described similar value, it is judged that Whether corresponding descreening facial image belongs to same people with Initial Face image.

First determines subelement 3043, if for judging to belong to same people, it is determined that face verification is passed through； Second determines subelement 3044, if for judging to be not belonging to same people, it is determined that face verification is not passed through.

Wherein, described judgment sub-unit 3042 can be specifically for: if described similar value exceedes predetermined threshold value, Then determine that corresponding descreening facial image and Initial Face image belong to same people, if described similar value does not surpasses Cross predetermined threshold value, it is determined that corresponding descreening facial image is not belonging to same people with Initial Face image.

In some embodiments, described computation subunit 3041 can be specifically for: utilizes Euclidean distance to calculate Any one in method, COS distance algorithm, associating bayesian algorithm and metric learning algorithm, goes described Similarity between reticulate pattern face characteristic and described Initial Face feature calculates.

When being embodied as, above unit can realize as independent entity, it is also possible to carries out arbitrarily Combination, realizes as same or several entities, and being embodied as of above unit can be found in above Embodiment of the method, does not repeats them here.

This face verification device specifically can be integrated in the network equipments such as server.

From the foregoing, the face verification device that the embodiment of the present invention provides, first, utilize that presets to remove net Stricture of vagina model carries out descreening operation to the reticulate pattern facial image needing checking automatically, obtains descreening facial image； Then, then by default descreening face verification model to descreening facial image and Initial Face image Carry out feature extraction respectively, obtain descreening face characteristic and Initial Face feature；It is finally based on descreening Face characteristic and Initial Face feature, carry out face to corresponding descreening facial image and Initial Face image Checking；I.e. the embodiment of the present invention is according to the good descreening model of training in advance and descreening face verification model, The reticulate pattern facial image and Initial Face image needing checking is carried out automatic Treatment Analysis, it is not necessary to specially Industry personnel's manual operation, can automatically determine reticulate pattern facial image and whether Initial Face image is same People, improves efficiency and the accuracy of face verification.

4th embodiment

The embodiment of the present invention also provides for a kind of server, wherein can be with the face verification of the integrated embodiment of the present invention Device, as shown in Figure 4, it illustrates the structural representation of server involved by the embodiment of the present invention, tool From the point of view of body:

This server can include one or the processor 401, or of more than one process core The memorizer 402 of above computer-readable recording medium, radio frequency (Radio Frequency, RF) circuit 403, The parts such as power supply 404, input block 405 and display unit 406.It will be understood by those skilled in the art that Server architecture shown in Fig. 4 is not intended that the restriction to server, can include more more or more than diagram Few parts, or combine some parts, or different parts are arranged.Wherein:

Processor 401 is the control centre of this server, utilizes various interface and the whole server of connection Various piece, by run or perform be stored in the software program in memorizer 402 and/or module, and Call the data being stored in memorizer 402, perform the various functions of server and process data, thus right Server carries out integral monitoring.Optionally, processor 401 can include one or more process core；Preferably , processor 401 can integrated application processor and modem processor, wherein, application processor is main Processing operating system, user interface and application program etc., modem processor mainly processes radio communication. It is understood that above-mentioned modem processor can not also be integrated in processor 401.

Memorizer 402 can be used for storing software program and module, and processor 401 is stored in by operation The software program of reservoir 402 and module, thus perform the application of various function and data process.Memorizer 402 can mainly include store program area and storage data field, wherein, storage program area can store operating system, Application program (such as sound-playing function, image player function etc.) etc. needed at least one function；Deposit Storage data field can store the data etc. that the use according to server is created.Additionally, memorizer 402 can wrap Include high-speed random access memory, it is also possible to include nonvolatile memory, for example, at least one disk storage Device, flush memory device or other volatile solid-state parts.Correspondingly, memorizer 402 can also wrap Include Memory Controller, to provide the processor 401 access to memorizer 402.

During RF circuit 403 can be used for receiving and sending messages, the reception of signal and transmission, especially, by base station Downlink information receive after, transfer to one or more than one processor 401 process；It addition, will relate to The data of row are sent to base station.Generally, RF circuit 403 include but not limited to antenna, at least one amplifier, Tuner, one or more agitator, subscriber identity module (SIM) card, transceiver, bonder, Low-noise amplifier (LNA, Low Noise Amplifier), duplexer etc..Additionally, RF circuit 403 Can also be communicated with network and other equipment by radio communication.Described radio communication can use arbitrary communication Standard or agreement, include but not limited to global system for mobile communications (GSM, Global System of Mobile Communication), general packet radio service (GPRS, General Packet Radio Service), CDMA (CDMA, Code Division Multiple Access), WCDMA (WCDMA, Wideband Code Division Multiple Access), Long Term Evolution (LTE, Long Term Evolution), Email, Short Message Service (SMS, Short Messaging Service) etc..

Server also includes the power supply 404 (such as battery) powered to all parts, it is preferred that power supply can With logically contiguous with processor 401 by power-supply management system, thus realize management by power-supply management system The functions such as charging, electric discharge and power managed.Power supply 404 can also include one or more directly Stream or alternating current power supply, recharging system, power failure detection circuit, power supply changeover device or inverter, electricity The random component such as source positioning indicator.

This server may also include input block 405, and this input block 405 can be used for receiving the numeral of input Or character information, and produce the keyboard relevant with user setup and function control, mouse, action bars, Optics or the input of trace ball signal.

This server may also include display unit 406, and this display unit 406 can be used for display and inputted by user Information or be supplied to the information of user and the various graphical user interface of server, these graphical users connect Mouth can be made up of figure, text, icon, video and its combination in any.Display unit 406 can include Display floater, optionally, can use liquid crystal display (LCD, Liquid Crystal Display), The forms such as Organic Light Emitting Diode (OLED, Organic Light-Emitting Diode) configure display surface Plate.

Concrete the most in the present embodiment, the processor 401 in server can according to following instruction, by one or The executable file that the process of more than one application program is corresponding is loaded in memorizer 402, and by processing Device 401 runs the application program being stored in memorizer 402, thus realizes various function, as follows:

Reticulate pattern facial image that acquisition need to be verified and Initial Face image；Utilize the descreening model pair preset Described reticulate pattern facial image carries out descreening operation, obtains descreening facial image；By default descreening Face verification model carries out feature extraction to described descreening facial image, obtains descreening face characteristic, and By described default descreening face verification model, described Initial Face image is carried out feature extraction, obtain Initial Face feature；Based on described descreening face characteristic and described Initial Face feature, to removing net accordingly Stricture of vagina facial image and Initial Face image carry out face verification.

Preferably, described processor 401 can be also used for: collects first kind facial image, and according to described First kind facial image builds the first face data set；Based on default reticulate pattern watermark, to described first face number Synthesize according to collection, obtain the first reticulate pattern human face data collection；Utilize described first face data set and described One reticulate pattern human face data collection carries out descreening convolutional neural networks training, generates descreening model；Go described Reticulate pattern model is defined as the descreening model preset.

Preferably, described processor 401 can be also used for: collects Equations of The Second Kind facial image, and according to described Equations of The Second Kind facial image builds the second human face data collection；Obtain the 3rd class that described second human face data collection is corresponding Facial image, and build the 3rd human face data collection according to described 3rd class facial image；Utilize described second people Face data set and described 3rd human face data collection carry out face verification convolutional neural networks training, generate the first Face checking model.

Preferably, described processor 401 can be also used for: based on default reticulate pattern watermark, to described second people Face data set synthesizes, and obtains the second reticulate pattern human face data collection；Utilize described descreening model to described Every image that two reticulate pattern human face data are concentrated carries out descreening operation, generates descreening human face data collection；Root According to described descreening human face data collection and described 3rd human face data collection, described first face verification model is entered Row network is finely tuned, and generates the second face verification model；It is defined as described second face verification model presetting Descreening face verification model.

Preferably, described processor 401 can be also used for: to described descreening face characteristic and described initially Similarity between face characteristic calculates, and obtains similar value；Based on described similar value, it is judged that corresponding Whether descreening facial image and Initial Face image belong to same people；If judging to belong to same people, the most really Determine face verification to pass through；If judging to be not belonging to same people, it is determined that face verification is not passed through.

Preferably, described processor 401 can be also used for: if described similar value exceedes predetermined threshold value, the most really Fixed corresponding descreening facial image belongs to same people with Initial Face image；If described similar value is less than pre- If threshold value, it is determined that corresponding descreening facial image is not belonging to same people with Initial Face image.

Preferably, described processor 401 can be also used for: utilize Euclidean distance algorithm, COS distance algorithm, Any one in associating bayesian algorithm and metric learning algorithm, to described descreening face characteristic and described Similarity between Initial Face feature calculates.

From the foregoing, in the server of embodiment of the present invention offer, first, utilize the descreening mould preset Type carries out descreening operation to the reticulate pattern facial image needing checking automatically, obtains descreening facial image；So After, then by the descreening face verification model preset, descreening facial image and Initial Face image are divided Do not carry out feature extraction, obtain descreening face characteristic and Initial Face feature；It is finally based on descreening people Face feature and Initial Face feature, carry out face to corresponding descreening facial image and Initial Face image and test Card；I.e. the embodiment of the present invention is according to the good descreening model of training in advance and descreening face verification model, The reticulate pattern facial image and Initial Face image needing checking is carried out automatic Treatment Analysis, it is not necessary to specially Industry personnel's manual operation, can automatically determine reticulate pattern facial image and whether Initial Face image is same People, improves efficiency and the accuracy of face verification.

In the above-described embodiments, the description to each embodiment all emphasizes particularly on different fields, in certain embodiment the most in detail The part stated, may refer to the detailed description above with respect to face verification method, and here is omitted.

The described face verification device that the embodiment of the present invention provides, is such as computer, panel computer, has Mobile phone of touch function etc., described face verification device belongs to the face verification method in foregoing embodiments Same design, can run offer in described face verification embodiment of the method on described face verification device Either method, it implements process and refers to described face verification embodiment of the method, and here is omitted.

It should be noted that for face verification method of the present invention, this area common test personnel can To understand all or part of flow process of the face verification method described in the embodiment of the present invention that realizes, can be by meter The hardware that calculation machine program controls to be correlated with completes, and described computer program can be stored in an embodied on computer readable In storage medium, as being stored in the memorizer of terminal, and performed by least one processor in this terminal, The flow process of embodiment such as described face verification method can be included in the process of implementation.Wherein, described storage Medium can be magnetic disc, CD, read only memory (ROM, Read Only Memory), random access memory note Recall body (RAM, Random Access Memory) etc..

For the described face verification device of the embodiment of the present invention, its each functional module can be integrated in one Process in chip, it is also possible to be that modules is individually physically present, it is also possible to two or more module collection In Cheng Yi module.Above-mentioned integrated module both can realize to use the form of hardware, it would however also be possible to employ soft The form of part functional module realizes.If described integrated module realizes with the form of software function module and makees During for independent production marketing or use, it is also possible to be stored in a computer read/write memory medium, institute State storage medium such as read only memory, disk or CD etc..

A kind of face verification method and device provided the embodiment of the present invention above is described in detail, Principle and the embodiment of the present invention are set forth by specific case used herein, above example Method and the core concept thereof being only intended to help to understand the present invention is described；Technology simultaneously for this area Personnel, according to the thought of the present invention, the most all will change, and combine Upper described, this specification content should not be construed as limitation of the present invention.

Claims

1. a face verification method, it is characterised in that including:

Face verification method the most according to claim 1, it is characterised in that described acquisition reticulate pattern face Before image and Initial Face image, also include:

Collect first kind facial image, and build the first face data set according to described first kind facial image；

Based on default reticulate pattern watermark, described first face data set is synthesized, obtains the first reticulate pattern face Data set；

Described first face data set and described first reticulate pattern human face data collection is utilized to carry out descreening convolutional Neural Network training, generates descreening model；

Described descreening model is defined as the descreening model preset.

Face verification method the most according to claim 2, it is characterised in that described acquisition reticulate pattern face Before image and Initial Face image, also include:

Collect Equations of The Second Kind facial image, and build the second human face data collection according to described Equations of The Second Kind facial image；

Obtain the 3rd class facial image that described second human face data collection is corresponding, and according to described 3rd class face Picture construction the 3rd human face data collection；

Described second human face data collection and described 3rd human face data collection is utilized to carry out face verification convolutional Neural net Network training, generates the first face verification model.

Face verification method the most according to claim 3, it is characterised in that described acquisition reticulate pattern face Before image and Initial Face image, also include:

Based on default reticulate pattern watermark, described second human face data collection is synthesized, obtains the second reticulate pattern face Data set；

Every the image utilizing described descreening model to concentrate described second reticulate pattern human face data carries out descreening Operation, generates descreening human face data collection；

According to described descreening human face data collection and described 3rd human face data collection, to described first face verification Model carries out network fine setting, generates the second face verification model；

Described second face verification model is defined as the descreening face verification model preset.

Face verification method the most according to claim 1, it is characterised in that described remove net based on described Stricture of vagina face characteristic and described Initial Face feature, enter corresponding descreening facial image and Initial Face image Row face verification, including:

Similarity between described descreening face characteristic and described Initial Face feature is calculated, obtains Similar value；

Based on described similar value, it is judged that whether corresponding descreening facial image and Initial Face image belong to same One people；

If judging to belong to same people, it is determined that face verification is passed through；

If judging to be not belonging to same people, it is determined that face verification is not passed through.

Face verification method the most according to claim 5, it is characterised in that described based on described similar Value, it is judged that whether corresponding descreening facial image belongs to same people with Initial Face image, including:

If described similar value exceedes predetermined threshold value, it is determined that corresponding descreening facial image and Initial Face figure As belonging to same people；

If described similar value is less than predetermined threshold value, it is determined that corresponding descreening facial image and Initial Face Image is not belonging to same people.

Face verification method the most according to claim 5, it is characterised in that described to described descreening Similarity between face characteristic and described Initial Face feature calculates, including:

Utilize in Euclidean distance algorithm, COS distance algorithm, associating bayesian algorithm and metric learning algorithm Any one, the similarity between described descreening face characteristic and described Initial Face feature is calculated.

8. a face verification device, it is characterised in that including:

Face verification device the most according to claim 8, it is characterised in that described device also includes:

First construction unit, is used for collecting first kind facial image, and according to described first kind facial image structure Build the first face data set；

First synthesis unit, for based on default reticulate pattern watermark, synthesizes described first face data set, Obtain the first reticulate pattern human face data collection；

First training unit, is used for utilizing described first face data set and described first reticulate pattern human face data collection Carry out descreening convolutional neural networks training, generate descreening model；

First determines unit, for described descreening model is defined as the descreening model preset.

Face verification device the most according to claim 9, it is characterised in that described device also includes:

Second construction unit, is used for collecting Equations of The Second Kind facial image, and according to described Equations of The Second Kind facial image structure Build the second human face data collection；

3rd construction unit, for obtaining the 3rd class facial image that described second human face data collection is corresponding, and The 3rd human face data collection is built according to described 3rd class facial image；

Second training unit, is used for utilizing described second human face data collection and described 3rd human face data collection to carry out Face verification convolutional neural networks is trained, and generates the first face verification model.

11. face verification devices according to claim 10, it is characterised in that described device also includes:

Second synthesis unit, for based on default reticulate pattern watermark, synthesizes described second human face data collection, Obtain the second reticulate pattern human face data collection；

Second descreening unit, is used for utilizing described descreening model to concentrate described second reticulate pattern human face data Every image carry out descreening operation, generate descreening human face data collection；

Fine-adjusting unit, for according to described descreening human face data collection and described 3rd human face data collection, to institute State the first face verification model and carry out network fine setting, generate the second face verification model；

Second determines unit, for being tested by the descreening face that described second face verification model is defined as presetting Model of a syndrome.

12. face verification devices according to claim 8, it is characterised in that described authentication unit bag Include:

Computation subunit, similar between described descreening face characteristic and described Initial Face feature Degree calculates, and obtains similar value；

Judgment sub-unit, for based on described similar value, it is judged that corresponding descreening facial image and initial people Whether face image belongs to same people；

First determines subelement, if for judging to belong to same people, it is determined that face verification is passed through；

Second determines subelement, if for judging to be not belonging to same people, it is determined that face verification is not passed through.

13. face verification devices according to claim 12, it is characterised in that described judgment sub-unit For:

If described similar value exceedes predetermined threshold value, it is determined that corresponding descreening facial image and Initial Face figure As belonging to same people, if described similar value is less than predetermined threshold value, it is determined that corresponding descreening facial image It is not belonging to same people with Initial Face image.

14. face verification devices according to claim 12, it is characterised in that described computation subunit For: